Input-Centric AI Governance

Every commercial AI bill is an output bill. Platforms price the tokens a model returns, finance teams treat that invoice as the cost of AI, and the resulting figure is tidy, auditable, and wrong about where the money went.

Humans work on the other side. They assemble context, engineer prompts, set guardrails, and re-enter the pipeline when an output fails. The operator never controls temperature, top-p, or top-k — those are locked platform constants at the application layer. The input payload is the only lever available, which is why this research treats input as the operational variable and output as a telemetry point.

Three consequences follow. The output is the trigger that reveals whether an input payload succeeded. A flawed output volume dictates how much work must re-enter the pipeline. And the input, not the output, is where operator intent and prompt efficiency actually live.

How much an organization can control inference depends entirely on how it deploys the model. Only the third tier gives an enterprise real control over behavior; the first two leave the operator with a text box and a bill.

In tiers 1 and 2 the input payload is the only variable a person can move. Governance that measures the model instead of the payload is measuring the one thing the enterprise does not command.

Enterprise AI metacognitive myopia is the failure to notice that financial accounting and operational labor move in opposite directions during AI integration. It surfaces through a set of structural paradoxes — each defensible on its own, and each hiding the same cost.

Existing security postures fail to measure AI operational friction for a structural reason, not a coverage gap. Post-hoc telemetry settles the account after the labor has already been spent, and in-loop middleware shares the fate of the system it is meant to police.

Traditional SaaS maintenance declines after deployment: the vendor patches the code. Generative systems invert that curve. Bugs are not broken code — they are context drifts, attention dilutions, and hallucination consensus, and none of them can be patched by a vendor.

That is why the cost of an AI program can rise while the API bill stays flat, and why headcount absorbs it silently.

Organizations reach for AI-on-AI review — one agent generates, a second assesses — as a substitute for human separation of duties. It fails for two reasons.

The skin-in-the-game deficit. Human controls work because people carry asymmetric risk: a missed oversight ends in termination, loss, or liability. An agent carries none. When the loop fails, the organization absorbs the consequence.

The mandate collapse. Assign an agent a checker role and it still optimizes the same mandate as the maker: a successful completion sequence built on token predictability. Agents on the same foundation model share probability distributions and blind spots, so a plausible hallucination is validated rather than challenged. The result is hallucination consensus, not adversarial review.

A single human placed at the end of a long pipeline does not fix this. It produces automation bias, and the reviewer rubber-stamps an output described as audited. Human-in-the-loop has to mean continuous operational oversight — the judgment that processes reputation, legal constraint, and brand equity, which no token distribution contains.

Shadow labor is countable only if each hour lands in exactly one bucket. The lexicon defines three baseline pillars, with inclusion and exclusion rules that keep rework, verification, and drift from being counted twice.

Perimeter variables

Four additional variables describe the exposure around sanctioned use — the activity that never reaches an enterprise invoice at all.

The cost does not arrive as a new invoice. It is absorbed by people already on payroll, which is why it has to be valued at a fully burdened labor rate — salary, benefits, seat licenses, and payroll taxes — rather than at an hourly wage.

1. 1

   Gross margin compression

   No new check is written for an intervention. Fully burdened employees — salary, healthcare, SaaS seats, payroll taxes — pivot to fix the issue at the same loaded rate, and that pivot is not allocated into the cost of goods sold.

2. 2

   The CapEx versus OpEx inversion penalty

   Building AI capability is capitalized and amortized. Rework, verification, and drift maintenance are maintenance, which makes them operating expense — the opposite accounting treatment for work attached to the same system.

3. 3

   Job costing and allocation deficits

   Emergency tuning bills against the original project code, producing an immediate negative budget variance. Roadmaps slip because engineering and risk capacity were absorbed by undocumented maintenance.

4. 4

   Post-project variance analysis

   Controllers compare estimated fully burdened hours to realized burdened cost. SIBR is the number that shows how much realized capacity was eroded by reactive tuning, and SIR is the empirical baseline for forecasting the next project's shadow labor tax.

SIR states how much of the workforce the friction consumed. SIBR states what that cost in money. Both use hours logged inside a defined operational period, priced at FBLR for an individual or BBLR for a team.

Each example follows the same pattern: define the period's maximum capacity, scope the labor to one pillar only, and price the hours at the correct burdened rate. All three numbers below are the source papers' own modeling.

The four papers support the framework with the enterprise benchmarks below. They are the author's cited figures — treat them as directional evidence for the mechanism, not as audited statistics about your own organization.

• 58–60%Identity evasionShare of enterprise AI sessions originating from unmanaged personal accounts on corporate hardware.
• 400+ daysDwell surchargeAverage time high-risk shadow applications remain inside enterprise networks before remediation.
• 65% / 40%Exfiltration riskShare of data-leak incidents containing personally identifiable information, and containing intellectual property.
• 12–14 hLoop remediationUnbudgeted human hours per incident to restore a clean operational state after an agentic loop failure.
• 43%Regulatory surchargeOrganizations reporting AI inventory blindness, turning transparency obligations into fixed operational liabilities.

The papers also point to regulatory context: EU AI Act Article 50 transparency obligations and AI Office enforcement powers taking effect in August 2026, which turn inventory blindness from an IT gap into an operating liability. Confirm your own obligations with counsel.

The argument closes where it started. AI cannot be governed as standard, deterministic software without an unmonitored margin leak, and that leak compounds in the dark until someone measures the input side.

• Meter input-chain labor, not just output tokens — the payload is the only variable the operator controls.
• Scope every intervention hour to one pillar so rework, verification, and drift never double count.
• Price intervention at fully burdened rates and book it as OpEx; capitalizing it hides the variance.
• Publish a verifiable AI inventory before a regulator, an auditor, or a board asks for one.

The framework is published as the Shadow AI Burden Rate under an input-centric accounting standard. The next step is measuring it on your own workload.